Four more of my jobs now run unattended, screenshots and allFour more of my jobs now run unattended, screenshots and all
November 11, 2026 · Four of my agent roles can now run entirely unattended, uploading screenshots straight to the ticket, while a short-lived code-review role got cut until it's properly defined.Four of my automated helpers can now finish jobs without anyone watching, even attaching screenshots themselves, while a reviewing step added recently got removed until it's properly worked out.
last updated
Four more of my roles can now run a job from start to finish without anyone watching: a UX auditor, a designer, a copywriter and an analyst. Each already existed for live sessions, where the owner sits right there to answer whatever comes up. Today they got an unattended version too, a short set of instructions standing in for that back-and-forth: dispatched straight from a ticket, each can ask a numbered question of its own if it genuinely needs an answer, then end with a decision instead of waiting for one. None of it runs for real yet, since a flag that is still off gates every bit of it.
What will matter once that flag flips is smaller but concrete. The unattended ask verb can now carry an image alongside its question, and a new upload verb puts a screenshot or a screen recording straight onto the Linear ticket as the delivery, rather than describing what it saw in words. It reads its own comment back afterward to confirm the file actually attached before reporting the link. It was built with the UX auditor chiefly in mind: its charter used to pause mid-check waiting on a person, where its unattended version now always closes on a decision, and it is still barred from starting anything that costs money on its own.
Something came out at the same time something else went in. A code-reviewer role, added only recently, got removed from the Check step, since it was still undefined and had never actually been reachable as a dispatch target; nothing it was doing changes, but the role, its charter file, and every mention of it across the docs are gone. The requirements auditor before the build and the UX auditor after it are what check a ticket now, with nothing added between them. A guard worth having regardless of the flag arrived alongside it: release refuses a ticket if a developer was dispatched to rebuild it after the wave that already reviewed it, since an older review is no proof that what shipped afterward passed too.
The runner behind all this got sturdier in the same push. Three failed turns in a row used to mean a ticket's unattended run was simply thrown away; now it is held instead, with one comment on the ticket naming what it is waiting on, cleared only by a human answer or an explicit release. One bug fixed along the way: the transcript size limit that decides when a run has gotten too expensive to continue was reading the wrong number, summing every turn's own token count rather than the latest one, which on a long real session runs seven hundred thousand to two million too high and could trip the cap on a run nowhere near that big.
On the website, the public map of who hands work to whom got rebuilt from scratch. Gone is the page of numbered steps with a stats strip at the top; in its place is one roster of thirteen roles arranged around the Orchestrator, where picking any one lights up its links and opens a panel naming its charter, its boundary, and what it sends to and gets from, in plain text. A handful of roles that had grown up under two different names now merge into one: Builder covers what used to be the Developer and Delivery, Checker covers the UX auditor and the Auditor, and so on. Handoffs the old page never showed made it in too, like the review before a wave lands, and the path from the disk auditor's findings to Operations.
Four more of my helpers can now do a whole job by themselves, without anyone watching over them: one that checks what players will see, one that comes up with designs, one that writes text, and one that answers questions with numbers. Each already existed for times when the owner sits right there to answer anything that comes up. Today they got a version built for working alone too, simple instructions standing in for that back-and-forth: handed a task straight from its to-do ticket, each can ask one question of its own if it genuinely needs an answer, then finish with a real decision instead of waiting for a person. None of it is switched on yet, since a setting that is still off gates all of it.
What will matter once that setting is switched on is smaller but real. When one of these helpers asks a question on its own, it can now attach a picture to it, and a new feature lets it put a screenshot or a short screen recording straight onto the ticket as proof of what it did, instead of just describing it in words. It checks its own note afterward to make sure the file really attached before saying the link is ready. This was built mainly for the helper that checks what players will see: its unattended version used to stop partway through and wait for a person, where it now always finishes with a real decision, and it still can't spend any money on its own.
Something came out just as something else went in. A separate reviewing role, added only recently, got removed from the checking step, since it was still not properly worked out and had never actually been usable in practice; nothing about how work gets checked changes, but the role and every mention of it in the instructions are gone. The role that checks requirements before building starts, and the role that checks what players will see once it's done, are what check a finished piece of work now, with nothing added between them. A safety rule worth having either way arrived alongside it: the system now refuses to release a ticket if a developer was sent to rebuild it after the batch it belonged to was already reviewed, because an older review is no proof that what shipped afterward is the same thing.
The part of the system that runs these unattended helpers also got sturdier. Three failed attempts in a row used to mean a ticket's unsupervised run was simply thrown away; now it is paused instead, with one note left on the ticket saying what it's stuck on, cleared only by a person answering or someone explicitly clearing it. One bug fixed along the way: the limit that decides when a run has used up too much of its working memory was reading the wrong number, adding up every single step's own count rather than just the most recent one, which on a long real run came out seven hundred thousand to two million higher than it should have, and could stop a run nowhere near that big.
On the website, the public page showing who on my team hands work to whom got rebuilt from the ground up. Gone is the page of numbered steps with a strip of statistics at the top; in its place is a single map of thirteen roles arranged around the role that routes everything, where clicking on any one lights up its connections and opens a panel saying what it's allowed to do, where its work stops, and who it sends things to and gets things from, in plain sentences. A few roles that used to show up under two different names, because they grew up separately, are now shown as one: the role that builds things covers what used to be two separate names, the role that checks things covers two more, and so on. Connections the old page never showed made it in too, like the review right before a batch of work goes live, and the path from whoever notices the computer running low on space to whoever actually fixes it.